File size: 8,751 Bytes
5143557
 
 
 
 
eed1cab
5143557
 
 
 
 
 
 
 
 
791c076
 
 
5143557
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
791c076
5143557
 
 
 
 
 
 
 
 
791c076
 
 
 
 
 
 
5143557
791c076
5143557
 
 
791c076
5143557
791c076
 
 
5143557
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eed1cab
791c076
 
5143557
 
 
791c076
 
5143557
791c076
 
 
 
 
eed1cab
 
5143557
 
 
 
 
791c076
 
 
 
 
 
5143557
 
791c076
 
 
 
 
5143557
 
 
 
 
eed1cab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5143557
 
 
 
791c076
eed1cab
 
 
 
 
 
5143557
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eed1cab
 
 
 
 
 
5143557
 
791c076
 
 
 
 
 
 
 
 
 
 
5143557
 
791c076
5143557
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
"""Top-level benchmark orchestration for agent comparison runs."""

from __future__ import annotations

import os
import sys
from pathlib import Path

from dotenv import load_dotenv

from dataforge.bench.core import (
    AggregateBenchmarkResult,
    BenchmarkRunOutput,
    SeedBenchmarkResult,
    aggregate_seed_results,
    build_benchmark_metadata,
    build_seed_list,
    dataset_evidence_from_loaded,
    estimate_llm_calls,
    validate_estimated_calls,
    write_run_output,
)
from dataforge.bench.groq_client import GroqBenchClient
from dataforge.bench.methods import (
    run_heuristic_episode,
    run_llm_react_episode,
    run_llm_zeroshot_episode,
    run_random_episode,
)
from dataforge.datasets.real_world import load_real_world_dataset
from dataforge.datasets.registry import DATASET_REGISTRY

_SUPPORTED_METHODS = frozenset({"random", "heuristic", "llm_zeroshot", "llm_react"})


def _validate_inputs(methods: list[str], datasets: list[str]) -> None:
    """Validate user-selected methods and datasets."""
    unknown_methods = sorted(set(methods) - _SUPPORTED_METHODS)
    unknown_datasets = sorted(set(datasets) - set(DATASET_REGISTRY))
    if unknown_methods:
        raise ValueError(f"Unknown benchmark methods: {unknown_methods}")
    if unknown_datasets:
        raise ValueError(f"Unknown benchmark datasets: {unknown_datasets}")


def _reproduction_command(
    methods: list[str],
    datasets: list[str],
    *,
    seed_count: int,
    seed_list: list[int] | None,
) -> str:
    """Build the canonical command for reproducing a benchmark run."""
    command = (
        "dataforge bench "
        f"--methods {','.join(methods)} "
        f"--datasets {','.join(datasets)} "
        f"--seeds {seed_count}"
    )
    if seed_list is not None:
        command += f" --seed-list {','.join(str(seed) for seed in seed_list)}"
    return command


def _llm_skip_reason() -> str | None:
    """Return a skip reason when LLM methods cannot run."""
    provider = os.environ.get("DATAFORGE_LLM_PROVIDER", "").strip().lower()
    if provider != "groq":
        return "DATAFORGE_LLM_PROVIDER must be set to groq."
    if not os.environ.get("GROQ_API_KEY"):
        return "GROQ_API_KEY is not set."
    return None


def _skipped_result(
    *,
    method: str,
    dataset: str,
    seed: int,
    reason: str,
    reproduction_command: str,
) -> SeedBenchmarkResult:
    """Build a skipped seed result with a clear reason."""
    return SeedBenchmarkResult(
        method=method,
        dataset=dataset,
        seed=seed,
        status="skipped",
        skip_reason=reason,
        llm_calls=0,
        prompt_tokens=0,
        completion_tokens=0,
        quota_units=0.0,
        runtime_s=0.0,
        provider=None,
        model=None,
        warnings=["provider_unset"],
        reproduction_command=reproduction_command,
    )


def run_agent_comparison(
    *,
    methods: list[str],
    datasets: list[str],
    seeds: int,
    output_json: Path,
    really_run_big_bench: bool,
    cache_root: Path | None = None,
    reproduction_command: str | None = None,
    seed_list: list[int] | None = None,
    verify_dataset_hashes: bool = True,
) -> BenchmarkRunOutput:
    """Run the selected benchmark methods across real-world datasets."""
    load_dotenv()
    _validate_inputs(methods, datasets)
    resolved_seed_list = build_seed_list(seeds=seeds, seed_list=seed_list)

    estimated_calls = estimate_llm_calls(
        methods=methods,
        datasets=datasets,
        seeds=len(resolved_seed_list),
    )
    # Validate call budget before any client instantiation or dataset loads that could
    # trigger network access in tests with environment variables set.
    validate_estimated_calls(
        estimated_calls=estimated_calls,
        really_run_big_bench=really_run_big_bench,
    )

    reproduction_command = reproduction_command or _reproduction_command(
        methods,
        datasets,
        seed_count=len(resolved_seed_list),
        seed_list=seed_list,
    )
    records: list[SeedBenchmarkResult] = []
    loaded_datasets = {
        dataset_name: load_real_world_dataset(
            dataset_name,
            cache_root=cache_root,
            verify_hashes=verify_dataset_hashes,
        )
        for dataset_name in datasets
    }

    llm_methods_requested = any(method.startswith("llm_") for method in methods)
    skip_reason = _llm_skip_reason() if llm_methods_requested else None
    client = None
    if llm_methods_requested and skip_reason is None:
        # Allow env-driven tuning for tiny CI checks.
        model = os.environ.get("DATAFORGE_GROQ_MODEL", "llama-3.3-70b-versatile")
        try:
            min_interval_s = float(os.environ.get("DATAFORGE_GROQ_MIN_INTERVAL_S", "1.0"))
        except ValueError:
            min_interval_s = 1.0
        try:
            timeout_s = float(os.environ.get("DATAFORGE_GROQ_TIMEOUT_S", "30"))
        except ValueError:
            timeout_s = 30.0
        try:
            max_tokens = int(os.environ.get("DATAFORGE_GROQ_MAX_TOKENS", "256"))
        except ValueError:
            max_tokens = 256
        try:
            max_retries = int(os.environ.get("DATAFORGE_GROQ_MAX_RETRIES", "3"))
        except ValueError:
            max_retries = 3
        client = GroqBenchClient(
            api_key=os.environ["GROQ_API_KEY"],
            model=model,
            min_interval_s=min_interval_s,
            max_tokens=max_tokens,
            max_retries=max_retries,
            timeout_s=timeout_s,
        )

    for dataset_name in datasets:
        dataset = loaded_datasets[dataset_name]
        for method in methods:
            for seed in resolved_seed_list:
                if os.environ.get("DATAFORGE_BENCH_VERBOSE"):
                    print(
                        f"[dataforge bench] start method={method} dataset={dataset_name} seed={seed}",
                        file=sys.stderr,
                        flush=True,
                    )
                if method == "random":
                    result = run_random_episode(dataset, seed=seed)
                elif method == "heuristic":
                    result = run_heuristic_episode(dataset, seed=seed)
                elif method == "llm_zeroshot":
                    if client is None or skip_reason is not None:
                        result = _skipped_result(
                            method=method,
                            dataset=dataset_name,
                            seed=seed,
                            reason=skip_reason or "LLM client unavailable.",
                            reproduction_command=reproduction_command,
                        )
                    else:
                        result = run_llm_zeroshot_episode(dataset, seed=seed, client=client)
                else:
                    if client is None or skip_reason is not None:
                        result = _skipped_result(
                            method=method,
                            dataset=dataset_name,
                            seed=seed,
                            reason=skip_reason or "LLM client unavailable.",
                            reproduction_command=reproduction_command,
                        )
                    else:
                        result = run_llm_react_episode(dataset, seed=seed, client=client)
                if result.reproduction_command != reproduction_command:
                    result = result.model_copy(
                        update={"reproduction_command": reproduction_command}
                    )
                if method == "heuristic":
                    result = result.model_copy(update={"seed": seed})
                records.append(result)
                if os.environ.get("DATAFORGE_BENCH_VERBOSE"):
                    print(
                        f"[dataforge bench] done  method={method} dataset={dataset_name} seed={seed} status={result.status}",
                        file=sys.stderr,
                        flush=True,
                    )

    aggregates: list[AggregateBenchmarkResult] = aggregate_seed_results(
        records, seeds_requested=len(resolved_seed_list)
    )
    dataset_evidence = [
        dataset_evidence_from_loaded(loaded_datasets[dataset_name]) for dataset_name in datasets
    ]
    metadata = build_benchmark_metadata(
        methods=methods,
        datasets=datasets,
        seed_list=resolved_seed_list,
        reproduction_command=reproduction_command,
        dataset_evidence=dataset_evidence,
    )
    output = BenchmarkRunOutput(
        metadata=metadata.model_dump(mode="json"),
        records=records,
        aggregates=aggregates,
    )
    write_run_output(output, output_json)
    return output